chore: session auto-commit
This commit is contained in:
MythEclipse
@@ -1,37 +1,11 @@
|
||||
import OpenAI from "openai";
|
||||
import { config } from "../../shared/config/config.js";
|
||||
import { createChildLogger } from "../../shared/logger/logger.js";
|
||||
import { getCachedText, upsertCachedText } from "./textCacheStore.js";
|
||||
import { llmDetectBadwords } from "./llmClient.js";
|
||||
|
||||
const log = createChildLogger("indonesianTextNormalizer");
|
||||
|
||||
const CUSTOM_EMOJI_PATTERN = /<a?:([a-zA-Z0-9_]+):(\d+)>/g;
|
||||
|
||||
const VALID_PRIMARY_AI_FLAGS = new Set([
|
||||
"spam",
|
||||
"hate_speech",
|
||||
"sara",
|
||||
"hoaks",
|
||||
"harassment",
|
||||
"vulgar_language",
|
||||
"sexual_content",
|
||||
"sexual_deviation",
|
||||
"violence",
|
||||
"self_harm",
|
||||
"doxxing",
|
||||
"scam",
|
||||
"misinformation",
|
||||
"nsfw_image",
|
||||
"gore_image",
|
||||
"illegal_content",
|
||||
"gambling",
|
||||
"drugs",
|
||||
"child_safety",
|
||||
"financial_scam",
|
||||
"religious_insult",
|
||||
"self_promo",
|
||||
]);
|
||||
|
||||
/**
|
||||
* In-memory cache TTL (10 min) — fastest path for repeated identical texts.
|
||||
*/
|
||||
@@ -51,7 +25,6 @@ interface BadwordCacheEntry {
|
||||
|
||||
const badwordCache = new Map<string, BadwordCacheEntry>();
|
||||
const inFlightBadwordLookups = new Map<string, Promise<string[]>>();
|
||||
let primaryModerationClient: OpenAI | null = null;
|
||||
|
||||
export interface ModerationTextEvidence {
|
||||
raw: string;
|
||||
@@ -121,117 +94,6 @@ function setCachedBadwords(key: string, value: string[]): void {
|
||||
}
|
||||
}
|
||||
|
||||
function getPrimaryModerationClient(): OpenAI | null {
|
||||
if (!config.AI_LLM_API_KEY) {
|
||||
return null;
|
||||
}
|
||||
|
||||
if (!primaryModerationClient) {
|
||||
primaryModerationClient = new OpenAI({
|
||||
apiKey: config.AI_LLM_API_KEY,
|
||||
baseURL: config.AI_LLM_BASE_URL,
|
||||
maxRetries: 0,
|
||||
timeout: 15000,
|
||||
});
|
||||
}
|
||||
|
||||
return primaryModerationClient;
|
||||
}
|
||||
|
||||
function normalizePrimaryAiFlag(value: string): string | null {
|
||||
const lower = value
|
||||
.trim()
|
||||
.toLowerCase()
|
||||
.replace(/[\s-]+/g, "_");
|
||||
if (!lower) return null;
|
||||
|
||||
if (VALID_PRIMARY_AI_FLAGS.has(lower)) {
|
||||
return lower;
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
|
||||
function extractFlagsFromPrimaryAiContent(content: string): string[] {
|
||||
const flags = new Set<string>();
|
||||
let parsed: unknown;
|
||||
|
||||
try {
|
||||
parsed = JSON.parse(content);
|
||||
} catch {
|
||||
parsed = null;
|
||||
}
|
||||
|
||||
const addValue = (value: unknown) => {
|
||||
if (typeof value !== "string") return;
|
||||
const normalized = normalizePrimaryAiFlag(value);
|
||||
if (normalized) flags.add(normalized);
|
||||
};
|
||||
|
||||
if (Array.isArray(parsed)) {
|
||||
for (const item of parsed) {
|
||||
addValue(item);
|
||||
}
|
||||
} else if (parsed && typeof parsed === "object") {
|
||||
const candidate = parsed as Record<string, unknown>;
|
||||
for (const key of ["flags", "categories", "badwords"]) {
|
||||
const value = candidate[key];
|
||||
if (Array.isArray(value)) {
|
||||
for (const item of value) addValue(item);
|
||||
} else {
|
||||
addValue(value);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (flags.size > 0) {
|
||||
return Array.from(flags);
|
||||
}
|
||||
|
||||
const lowerContent = content.toLowerCase();
|
||||
for (const flag of VALID_PRIMARY_AI_FLAGS) {
|
||||
if (lowerContent.includes(flag)) {
|
||||
flags.add(flag);
|
||||
}
|
||||
}
|
||||
|
||||
return Array.from(flags);
|
||||
}
|
||||
|
||||
async function callPrimaryAiModeration(text: string): Promise<string[]> {
|
||||
const client = getPrimaryModerationClient();
|
||||
if (!client) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const completion = await client.chat.completions.create({
|
||||
model: config.AI_LLM_MODEL,
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content:
|
||||
"Deteksi kata kasar / pelanggaran ringan dari teks Indonesia berikut. " +
|
||||
'Balas hanya JSON object dengan format {"flags":[...]} dan gunakan hanya flag valid ini: ' +
|
||||
Array.from(VALID_PRIMARY_AI_FLAGS).join(", ") +
|
||||
". Jika tidak ada pelanggaran, flags harus array kosong. Teks: " +
|
||||
text,
|
||||
},
|
||||
],
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
max_tokens: 200,
|
||||
stream: false,
|
||||
response_format: { type: "json_object" },
|
||||
} as OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming);
|
||||
|
||||
const content = completion.choices[0]?.message?.content?.trim();
|
||||
if (!content) {
|
||||
return [];
|
||||
}
|
||||
|
||||
return extractFlagsFromPrimaryAiContent(content);
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Two-tier cache + Primary AI pipeline
|
||||
// ---------------------------------------------------------------------------
|
||||
@@ -244,7 +106,8 @@ async function callPrimaryAiModeration(text: string): Promise<string[]> {
|
||||
* 2. **DB cache** (DB_CACHE_TTL_MS, 24 h) — same full-text key, persisted
|
||||
* across restarts. Uses the FULL normalized text (not per-word) because
|
||||
* context matters: "kau" alone is clean, but "awas kau" can be a threat.
|
||||
* 3. **Primary AI** (AI_LLM endpoint) — only runs when both cache layers miss.
|
||||
* 3. **Primary AI** (AI_LLM endpoint via llmClient) — only runs when both
|
||||
* cache layers miss.
|
||||
*
|
||||
* No local hardcoded badword list — all detection goes through AI APIs
|
||||
* to eliminate false positives from substring matching.
|
||||
@@ -275,10 +138,10 @@ export async function detectIndonesianBadwords(
|
||||
return flags;
|
||||
}
|
||||
|
||||
// ── Tier 3: Primary AI only ──
|
||||
// ── Tier 3: Primary AI only (via centralized llmClient) ──
|
||||
let finalHits: string[] = [];
|
||||
try {
|
||||
finalHits = await callPrimaryAiModeration(text);
|
||||
finalHits = await llmDetectBadwords(text);
|
||||
} catch (error) {
|
||||
log.warn(
|
||||
{ error: error instanceof Error ? error.message : String(error) },
|
||||
|
||||
@@ -0,0 +1,276 @@
|
||||
/**
|
||||
* Centralised LLM chat completion helper.
|
||||
*
|
||||
* All `openai.chat.completions.create` calls in the moderation subsystem
|
||||
* go through this module so that model, concurrency, retry, and token
|
||||
* defaults are maintained in one place.
|
||||
*/
|
||||
|
||||
import OpenAI from "openai";
|
||||
import { config } from "../../shared/config/config.js";
|
||||
import { retryWithBackoff } from "../../shared/utils/retry.js";
|
||||
import { withLlmConcurrency } from "./concurrencyLimiter.js";
|
||||
import { createChildLogger } from "../../shared/logger/logger.js";
|
||||
|
||||
const log = createChildLogger("llm-client");
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Lazy singleton — created on first use so that config is always resolved.
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
let openaiClient: OpenAI | null = null;
|
||||
|
||||
function getClient(): OpenAI | null {
|
||||
if (!config.AI_LLM_API_KEY) return null;
|
||||
if (!openaiClient) {
|
||||
openaiClient = new OpenAI({
|
||||
apiKey: config.AI_LLM_API_KEY,
|
||||
baseURL: config.AI_LLM_BASE_URL,
|
||||
maxRetries: 0,
|
||||
timeout: 15_000,
|
||||
});
|
||||
}
|
||||
return openaiClient;
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Shared defaults
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const DEFAULT_TEMPERATURE = 0.2;
|
||||
const DEFAULT_TOP_P = 0.95;
|
||||
const DEFAULT_RETRIES = 2;
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Public API
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
export interface LlmCallOpts {
|
||||
/** Conversation to send. Either a string (→ single user message) or an array of messages. */
|
||||
messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[];
|
||||
/** Which model to use (defaults to config.AI_LLM_MODEL). */
|
||||
model?: string;
|
||||
/** Max output tokens (defaults to 8192). */
|
||||
max_tokens?: number;
|
||||
/** Temperature (defaults to 0.2). */
|
||||
temperature?: number;
|
||||
/** Top-p (defaults to 0.95). */
|
||||
top_p?: number;
|
||||
/** Force JSON output. When true, wraps schema in json_schema response_format. */
|
||||
jsonResponse?:
|
||||
| { type: "json_object" }
|
||||
| {
|
||||
type: "json_schema";
|
||||
name: string;
|
||||
schema: Record<string, unknown>;
|
||||
strict: boolean;
|
||||
};
|
||||
/** Extra retries beyond DEFAULT_RETRIES (default 2). */
|
||||
retries?: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Call the LLM with sensible defaults: concurrency cap, retry, model, tokens.
|
||||
*
|
||||
* Returns the raw OpenAI ChatCompletion so callers can inspect
|
||||
* `choices[0].message.content`, `finish_reason`, `usage`, etc.
|
||||
*/
|
||||
export async function llmChat(
|
||||
opts: LlmCallOpts,
|
||||
): Promise<OpenAI.Chat.Completions.ChatCompletion | null> {
|
||||
const client = getClient();
|
||||
if (!client) return null;
|
||||
|
||||
const {
|
||||
messages,
|
||||
model = config.AI_LLM_MODEL,
|
||||
max_tokens = 8192,
|
||||
temperature = DEFAULT_TEMPERATURE,
|
||||
top_p = DEFAULT_TOP_P,
|
||||
jsonResponse,
|
||||
retries = DEFAULT_RETRIES,
|
||||
} = opts;
|
||||
|
||||
const responseFormat:
|
||||
| { type: "json_object" }
|
||||
| {
|
||||
type: "json_schema";
|
||||
json_schema: {
|
||||
name: string;
|
||||
schema: Record<string, unknown>;
|
||||
strict: boolean;
|
||||
};
|
||||
}
|
||||
| undefined = jsonResponse
|
||||
? jsonResponse.type === "json_schema"
|
||||
? {
|
||||
type: "json_schema",
|
||||
json_schema: {
|
||||
name: jsonResponse.name,
|
||||
schema: jsonResponse.schema,
|
||||
strict: jsonResponse.strict,
|
||||
},
|
||||
}
|
||||
: jsonResponse
|
||||
: undefined;
|
||||
|
||||
const params: OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming =
|
||||
{
|
||||
model,
|
||||
messages,
|
||||
temperature,
|
||||
top_p,
|
||||
max_tokens,
|
||||
stream: false,
|
||||
...(responseFormat ? { response_format: responseFormat } : {}),
|
||||
} as OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming;
|
||||
|
||||
return retryWithBackoff(
|
||||
async () => {
|
||||
return withLlmConcurrency(async () =>
|
||||
client.chat.completions.create(params),
|
||||
);
|
||||
},
|
||||
{
|
||||
retries,
|
||||
minTimeout: 0,
|
||||
maxTimeout: 0,
|
||||
factor: 2,
|
||||
logger: log,
|
||||
},
|
||||
);
|
||||
}
|
||||
|
||||
/**
|
||||
* Convenience for the legacy text-only badword detection call in
|
||||
* `indonesianTextNormalizer`. Returns parsed flags or [].
|
||||
*/
|
||||
export async function llmDetectBadwords(text: string): Promise<string[]> {
|
||||
const completion = await llmChat({
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content:
|
||||
"Deteksi kata kasar / pelanggaran ringan dari teks Indonesia berikut. " +
|
||||
'Balas hanya JSON object dengan format {"flags":[...]} dan gunakan hanya flag valid ini: ' +
|
||||
Array.from(VALID_PRIMARY_AI_FLAGS).join(", ") +
|
||||
". Jika tidak ada pelanggaran, flags harus array kosong. Teks: " +
|
||||
text,
|
||||
},
|
||||
],
|
||||
max_tokens: 200,
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
jsonResponse: { type: "json_object" },
|
||||
retries: 2,
|
||||
});
|
||||
|
||||
if (!completion) return [];
|
||||
const content = completion.choices[0]?.message?.content?.trim();
|
||||
if (!content) return [];
|
||||
return extractFlagsFromContent(content);
|
||||
}
|
||||
|
||||
/**
|
||||
* Convenience for vision (image/sticker/emoji) analysis.
|
||||
* Returns the raw completion content (trimmed) or null.
|
||||
*/
|
||||
export async function llmVision(
|
||||
promptText: string,
|
||||
imageUrl: { url: string },
|
||||
): Promise<string | null> {
|
||||
const completion = await llmChat({
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content: [
|
||||
{ type: "text" as const, text: promptText },
|
||||
{ type: "image_url" as const, image_url: imageUrl },
|
||||
],
|
||||
},
|
||||
],
|
||||
model: config.AI_LLM_VISION_MODEL ?? config.AI_LLM_MODEL,
|
||||
max_tokens: 500,
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
retries: 2,
|
||||
});
|
||||
|
||||
if (!completion) return null;
|
||||
return completion.choices[0]?.message?.content?.trim() ?? null;
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Flag extraction (reused from indonesianTextNormalizer)
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const VALID_PRIMARY_AI_FLAGS = new Set([
|
||||
"spam",
|
||||
"hate_speech",
|
||||
"sara",
|
||||
"hoaks",
|
||||
"harassment",
|
||||
"vulgar_language",
|
||||
"sexual_content",
|
||||
"sexual_deviation",
|
||||
"violence",
|
||||
"self_harm",
|
||||
"doxxing",
|
||||
"scam",
|
||||
"misinformation",
|
||||
"nsfw_image",
|
||||
"gore_image",
|
||||
"illegal_content",
|
||||
"gambling",
|
||||
"drugs",
|
||||
"child_safety",
|
||||
"financial_scam",
|
||||
"religious_insult",
|
||||
"self_promo",
|
||||
]);
|
||||
|
||||
function normalizeFlag(value: string): string | null {
|
||||
const lower = value.trim().toLowerCase().replace(/[\s-]+/g, "_");
|
||||
if (!lower) return null;
|
||||
if (VALID_PRIMARY_AI_FLAGS.has(lower)) return lower;
|
||||
return null;
|
||||
}
|
||||
|
||||
function extractFlagsFromContent(content: string): string[] {
|
||||
const flags = new Set<string>();
|
||||
let parsed: unknown;
|
||||
try {
|
||||
parsed = JSON.parse(content);
|
||||
} catch {
|
||||
parsed = null;
|
||||
}
|
||||
|
||||
const addValue = (v: unknown) => {
|
||||
if (typeof v !== "string") return;
|
||||
const n = normalizeFlag(v);
|
||||
if (n) flags.add(n);
|
||||
};
|
||||
|
||||
if (Array.isArray(parsed)) {
|
||||
for (const item of parsed) addValue(item);
|
||||
} else if (parsed && typeof parsed === "object") {
|
||||
const obj = parsed as Record<string, unknown>;
|
||||
for (const key of ["flags", "categories", "badwords"]) {
|
||||
const val = obj[key];
|
||||
if (Array.isArray(val)) {
|
||||
for (const item of val) addValue(item);
|
||||
} else {
|
||||
addValue(val);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (flags.size > 0) return Array.from(flags);
|
||||
|
||||
const lower = content.toLowerCase();
|
||||
for (const flag of VALID_PRIMARY_AI_FLAGS) {
|
||||
if (lower.includes(flag)) flags.add(flag);
|
||||
}
|
||||
|
||||
return Array.from(flags);
|
||||
}
|
||||
@@ -11,8 +11,8 @@ import type {
|
||||
AttachmentRecord,
|
||||
MessageRecord,
|
||||
} from "../message-capture/types.js";
|
||||
import { withLlmConcurrency } from "./concurrencyLimiter.js";
|
||||
import { formatModerationTextEvidenceForPrompt } from "./indonesianTextNormalizer.js";
|
||||
import { llmChat, llmVision } from "./llmClient.js";
|
||||
import { buildSystemPrompt as buildSystemPromptModular } from "./moderationPrompt.js";
|
||||
import { logModerationAnalysis, logModerationError } from "./responseLogger.js";
|
||||
import {
|
||||
@@ -592,39 +592,7 @@ const analyzeSingleMediaImage = async (
|
||||
: buildGeneralImageVisionPrompt(image.sourceLabel, messageId);
|
||||
|
||||
try {
|
||||
const completion = await retryWithBackoff(
|
||||
async () => {
|
||||
return withLlmConcurrency(async () =>
|
||||
openai.chat.completions.create({
|
||||
model: config.AI_LLM_VISION_MODEL ?? config.AI_LLM_MODEL,
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content: [
|
||||
{
|
||||
type: "text",
|
||||
text: promptText,
|
||||
},
|
||||
{ type: "image_url", image_url: image.image_url },
|
||||
],
|
||||
},
|
||||
],
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
max_tokens: 500,
|
||||
stream: false,
|
||||
} as OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming),
|
||||
);
|
||||
},
|
||||
{
|
||||
retries: 2,
|
||||
minTimeout: 0,
|
||||
maxTimeout: 0,
|
||||
logger: log,
|
||||
},
|
||||
);
|
||||
|
||||
const content = completion.choices[0]?.message?.content?.trim();
|
||||
const content = await llmVision(promptText, image.image_url);
|
||||
if (!content) return null;
|
||||
|
||||
await upsertCachedMediaAnalysis(
|
||||
@@ -693,26 +661,21 @@ async function callModerationLLM(
|
||||
try {
|
||||
const content = await buildContent(state);
|
||||
|
||||
const completion = await withLlmConcurrency(async () =>
|
||||
openai.chat.completions.create({
|
||||
model: config.AI_LLM_MODEL,
|
||||
messages: [{ role: "user", content }],
|
||||
temperature: 0.2,
|
||||
top_p: 0.95,
|
||||
// Sufficient for 20 moderation results (each ~70-150 tokens).
|
||||
// Previous 4096 caused LLM truncation after ~6 results.
|
||||
max_tokens: 16384,
|
||||
response_format: {
|
||||
type: "json_schema",
|
||||
json_schema: {
|
||||
name: "moderation_result",
|
||||
schema: MODERATION_JSON_SCHEMA,
|
||||
strict: true,
|
||||
},
|
||||
},
|
||||
stream: false,
|
||||
} as OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming),
|
||||
);
|
||||
const completion = await llmChat({
|
||||
messages: [{ role: "user", content }],
|
||||
max_tokens: 16384,
|
||||
jsonResponse: {
|
||||
type: "json_schema",
|
||||
name: "moderation_result",
|
||||
schema: MODERATION_JSON_SCHEMA,
|
||||
strict: true,
|
||||
},
|
||||
retries: 0,
|
||||
});
|
||||
|
||||
if (!completion) {
|
||||
throw new Error("LLM client unavailable (no API key)");
|
||||
}
|
||||
|
||||
if (
|
||||
!completion.choices ||
|
||||
|
||||
@@ -1,37 +1,11 @@
|
||||
import OpenAI from "openai";
|
||||
import { config } from "../config.js";
|
||||
import { createChildLogger } from "../logger.js";
|
||||
import { getCachedText, upsertCachedText } from "./textCacheStore.js";
|
||||
import { llmDetectBadwords } from "./llmClient.js";
|
||||
|
||||
const log = createChildLogger("indonesianTextNormalizer");
|
||||
|
||||
const CUSTOM_EMOJI_PATTERN = /<a?:([a-zA-Z0-9_]+):(\d+)>/g;
|
||||
|
||||
const VALID_PRIMARY_AI_FLAGS = new Set([
|
||||
"spam",
|
||||
"hate_speech",
|
||||
"sara",
|
||||
"hoaks",
|
||||
"harassment",
|
||||
"vulgar_language",
|
||||
"sexual_content",
|
||||
"sexual_deviation",
|
||||
"violence",
|
||||
"self_harm",
|
||||
"doxxing",
|
||||
"scam",
|
||||
"misinformation",
|
||||
"nsfw_image",
|
||||
"gore_image",
|
||||
"illegal_content",
|
||||
"gambling",
|
||||
"drugs",
|
||||
"child_safety",
|
||||
"financial_scam",
|
||||
"religious_insult",
|
||||
"self_promo",
|
||||
]);
|
||||
|
||||
/**
|
||||
* In-memory cache TTL (10 min) — fastest path for repeated identical texts.
|
||||
*/
|
||||
@@ -51,7 +25,6 @@ interface BadwordCacheEntry {
|
||||
|
||||
const badwordCache = new Map<string, BadwordCacheEntry>();
|
||||
const inFlightBadwordLookups = new Map<string, Promise<string[]>>();
|
||||
let primaryModerationClient: OpenAI | null = null;
|
||||
|
||||
export interface ModerationTextEvidence {
|
||||
raw: string;
|
||||
@@ -121,117 +94,6 @@ function setCachedBadwords(key: string, value: string[]): void {
|
||||
}
|
||||
}
|
||||
|
||||
function getPrimaryModerationClient(): OpenAI | null {
|
||||
if (!config.AI_LLM_API_KEY) {
|
||||
return null;
|
||||
}
|
||||
|
||||
if (!primaryModerationClient) {
|
||||
primaryModerationClient = new OpenAI({
|
||||
apiKey: config.AI_LLM_API_KEY,
|
||||
baseURL: config.AI_LLM_BASE_URL,
|
||||
maxRetries: 0,
|
||||
timeout: 15000,
|
||||
});
|
||||
}
|
||||
|
||||
return primaryModerationClient;
|
||||
}
|
||||
|
||||
function normalizePrimaryAiFlag(value: string): string | null {
|
||||
const lower = value
|
||||
.trim()
|
||||
.toLowerCase()
|
||||
.replace(/[\s-]+/g, "_");
|
||||
if (!lower) return null;
|
||||
|
||||
if (VALID_PRIMARY_AI_FLAGS.has(lower)) {
|
||||
return lower;
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
|
||||
function extractFlagsFromPrimaryAiContent(content: string): string[] {
|
||||
const flags = new Set<string>();
|
||||
let parsed: unknown;
|
||||
|
||||
try {
|
||||
parsed = JSON.parse(content);
|
||||
} catch {
|
||||
parsed = null;
|
||||
}
|
||||
|
||||
const addValue = (value: unknown) => {
|
||||
if (typeof value !== "string") return;
|
||||
const normalized = normalizePrimaryAiFlag(value);
|
||||
if (normalized) flags.add(normalized);
|
||||
};
|
||||
|
||||
if (Array.isArray(parsed)) {
|
||||
for (const item of parsed) {
|
||||
addValue(item);
|
||||
}
|
||||
} else if (parsed && typeof parsed === "object") {
|
||||
const candidate = parsed as Record<string, unknown>;
|
||||
for (const key of ["flags", "categories", "badwords"]) {
|
||||
const value = candidate[key];
|
||||
if (Array.isArray(value)) {
|
||||
for (const item of value) addValue(item);
|
||||
} else {
|
||||
addValue(value);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (flags.size > 0) {
|
||||
return Array.from(flags);
|
||||
}
|
||||
|
||||
const lowerContent = content.toLowerCase();
|
||||
for (const flag of VALID_PRIMARY_AI_FLAGS) {
|
||||
if (lowerContent.includes(flag)) {
|
||||
flags.add(flag);
|
||||
}
|
||||
}
|
||||
|
||||
return Array.from(flags);
|
||||
}
|
||||
|
||||
async function callPrimaryAiModeration(text: string): Promise<string[]> {
|
||||
const client = getPrimaryModerationClient();
|
||||
if (!client) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const completion = await client.chat.completions.create({
|
||||
model: config.AI_LLM_MODEL,
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content:
|
||||
"Deteksi kata kasar / pelanggaran ringan dari teks Indonesia berikut. " +
|
||||
'Balas hanya JSON object dengan format {"flags":[...]} dan gunakan hanya flag valid ini: ' +
|
||||
Array.from(VALID_PRIMARY_AI_FLAGS).join(", ") +
|
||||
". Jika tidak ada pelanggaran, flags harus array kosong. Teks: " +
|
||||
text,
|
||||
},
|
||||
],
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
max_tokens: 200,
|
||||
stream: false,
|
||||
response_format: { type: "json_object" },
|
||||
} as OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming);
|
||||
|
||||
const content = completion.choices[0]?.message?.content?.trim();
|
||||
if (!content) {
|
||||
return [];
|
||||
}
|
||||
|
||||
return extractFlagsFromPrimaryAiContent(content);
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Two-tier cache + Primary AI pipeline
|
||||
// ---------------------------------------------------------------------------
|
||||
@@ -244,7 +106,8 @@ async function callPrimaryAiModeration(text: string): Promise<string[]> {
|
||||
* 2. **DB cache** (DB_CACHE_TTL_MS, 24 h) — same full-text key, persisted
|
||||
* across restarts. Uses the FULL normalized text (not per-word) because
|
||||
* context matters: "kau" alone is clean, but "awas kau" can be a threat.
|
||||
* 3. **Primary AI** (AI_LLM endpoint) — only runs when both cache layers miss.
|
||||
* 3. **Primary AI** (AI_LLM endpoint via llmClient) — only runs when both
|
||||
* cache layers miss.
|
||||
*
|
||||
* No local hardcoded badword list — all detection goes through AI APIs
|
||||
* to eliminate false positives from substring matching.
|
||||
@@ -275,10 +138,10 @@ export async function detectIndonesianBadwords(
|
||||
return flags;
|
||||
}
|
||||
|
||||
// ── Tier 3: Primary AI only ──
|
||||
// ── Tier 3: Primary AI only (via centralized llmClient) ──
|
||||
let finalHits: string[] = [];
|
||||
try {
|
||||
finalHits = await callPrimaryAiModeration(text);
|
||||
finalHits = await llmDetectBadwords(text);
|
||||
} catch (error) {
|
||||
log.warn(
|
||||
{ error: error instanceof Error ? error.message : String(error) },
|
||||
|
||||
@@ -0,0 +1,276 @@
|
||||
/**
|
||||
* Centralised LLM chat completion helper.
|
||||
*
|
||||
* All `openai.chat.completions.create` calls in the moderation subsystem
|
||||
* go through this module so that model, concurrency, retry, and token
|
||||
* defaults are maintained in one place.
|
||||
*/
|
||||
|
||||
import OpenAI from "openai";
|
||||
import { config } from "../config.js";
|
||||
import { retryWithBackoff } from "../retry.js";
|
||||
import { withLlmConcurrency } from "./concurrencyLimiter.js";
|
||||
import { createChildLogger } from "../logger.js";
|
||||
|
||||
const log = createChildLogger("llm-client");
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Lazy singleton — created on first use so that config is always resolved.
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
let openaiClient: OpenAI | null = null;
|
||||
|
||||
function getClient(): OpenAI | null {
|
||||
if (!config.AI_LLM_API_KEY) return null;
|
||||
if (!openaiClient) {
|
||||
openaiClient = new OpenAI({
|
||||
apiKey: config.AI_LLM_API_KEY,
|
||||
baseURL: config.AI_LLM_BASE_URL,
|
||||
maxRetries: 0,
|
||||
timeout: 15_000,
|
||||
});
|
||||
}
|
||||
return openaiClient;
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Shared defaults
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const DEFAULT_TEMPERATURE = 0.2;
|
||||
const DEFAULT_TOP_P = 0.95;
|
||||
const DEFAULT_RETRIES = 2;
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Public API
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
export interface LlmCallOpts {
|
||||
/** Conversation to send. Either a string (→ single user message) or an array of messages. */
|
||||
messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[];
|
||||
/** Which model to use (defaults to config.AI_LLM_MODEL). */
|
||||
model?: string;
|
||||
/** Max output tokens (defaults to 8192). */
|
||||
max_tokens?: number;
|
||||
/** Temperature (defaults to 0.2). */
|
||||
temperature?: number;
|
||||
/** Top-p (defaults to 0.95). */
|
||||
top_p?: number;
|
||||
/** Force JSON output. When true, wraps schema in json_schema response_format. */
|
||||
jsonResponse?:
|
||||
| { type: "json_object" }
|
||||
| {
|
||||
type: "json_schema";
|
||||
name: string;
|
||||
schema: Record<string, unknown>;
|
||||
strict: boolean;
|
||||
};
|
||||
/** Extra retries beyond DEFAULT_RETRIES (default 2). */
|
||||
retries?: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Call the LLM with sensible defaults: concurrency cap, retry, model, tokens.
|
||||
*
|
||||
* Returns the raw OpenAI ChatCompletion so callers can inspect
|
||||
* `choices[0].message.content`, `finish_reason`, `usage`, etc.
|
||||
*/
|
||||
export async function llmChat(
|
||||
opts: LlmCallOpts,
|
||||
): Promise<OpenAI.Chat.Completions.ChatCompletion | null> {
|
||||
const client = getClient();
|
||||
if (!client) return null;
|
||||
|
||||
const {
|
||||
messages,
|
||||
model = config.AI_LLM_MODEL,
|
||||
max_tokens = 8192,
|
||||
temperature = DEFAULT_TEMPERATURE,
|
||||
top_p = DEFAULT_TOP_P,
|
||||
jsonResponse,
|
||||
retries = DEFAULT_RETRIES,
|
||||
} = opts;
|
||||
|
||||
const responseFormat:
|
||||
| { type: "json_object" }
|
||||
| {
|
||||
type: "json_schema";
|
||||
json_schema: {
|
||||
name: string;
|
||||
schema: Record<string, unknown>;
|
||||
strict: boolean;
|
||||
};
|
||||
}
|
||||
| undefined = jsonResponse
|
||||
? jsonResponse.type === "json_schema"
|
||||
? {
|
||||
type: "json_schema",
|
||||
json_schema: {
|
||||
name: jsonResponse.name,
|
||||
schema: jsonResponse.schema,
|
||||
strict: jsonResponse.strict,
|
||||
},
|
||||
}
|
||||
: jsonResponse
|
||||
: undefined;
|
||||
|
||||
const params: OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming =
|
||||
{
|
||||
model,
|
||||
messages,
|
||||
temperature,
|
||||
top_p,
|
||||
max_tokens,
|
||||
stream: false,
|
||||
...(responseFormat ? { response_format: responseFormat } : {}),
|
||||
} as OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming;
|
||||
|
||||
return retryWithBackoff(
|
||||
async () => {
|
||||
return withLlmConcurrency(async () =>
|
||||
client.chat.completions.create(params),
|
||||
);
|
||||
},
|
||||
{
|
||||
retries,
|
||||
minTimeout: 0,
|
||||
maxTimeout: 0,
|
||||
factor: 2,
|
||||
logger: log,
|
||||
},
|
||||
);
|
||||
}
|
||||
|
||||
/**
|
||||
* Convenience for the legacy text-only badword detection call in
|
||||
* `indonesianTextNormalizer`. Returns parsed flags or [].
|
||||
*/
|
||||
export async function llmDetectBadwords(text: string): Promise<string[]> {
|
||||
const completion = await llmChat({
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content:
|
||||
"Deteksi kata kasar / pelanggaran ringan dari teks Indonesia berikut. " +
|
||||
'Balas hanya JSON object dengan format {"flags":[...]} dan gunakan hanya flag valid ini: ' +
|
||||
Array.from(VALID_PRIMARY_AI_FLAGS).join(", ") +
|
||||
". Jika tidak ada pelanggaran, flags harus array kosong. Teks: " +
|
||||
text,
|
||||
},
|
||||
],
|
||||
max_tokens: 200,
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
jsonResponse: { type: "json_object" },
|
||||
retries: 2,
|
||||
});
|
||||
|
||||
if (!completion) return [];
|
||||
const content = completion.choices[0]?.message?.content?.trim();
|
||||
if (!content) return [];
|
||||
return extractFlagsFromContent(content);
|
||||
}
|
||||
|
||||
/**
|
||||
* Convenience for vision (image/sticker/emoji) analysis.
|
||||
* Returns the raw completion content (trimmed) or null.
|
||||
*/
|
||||
export async function llmVision(
|
||||
promptText: string,
|
||||
imageUrl: { url: string },
|
||||
): Promise<string | null> {
|
||||
const completion = await llmChat({
|
||||
messages: [
|
||||
{
|
||||
role: "user",
|
||||
content: [
|
||||
{ type: "text" as const, text: promptText },
|
||||
{ type: "image_url" as const, image_url: imageUrl },
|
||||
],
|
||||
},
|
||||
],
|
||||
model: config.AI_LLM_VISION_MODEL ?? config.AI_LLM_MODEL,
|
||||
max_tokens: 500,
|
||||
temperature: 0.1,
|
||||
top_p: 0.9,
|
||||
retries: 2,
|
||||
});
|
||||
|
||||
if (!completion) return null;
|
||||
return completion.choices[0]?.message?.content?.trim() ?? null;
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Flag extraction (reused from indonesianTextNormalizer)
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const VALID_PRIMARY_AI_FLAGS = new Set([
|
||||
"spam",
|
||||
"hate_speech",
|
||||
"sara",
|
||||
"hoaks",
|
||||
"harassment",
|
||||
"vulgar_language",
|
||||
"sexual_content",
|
||||
"sexual_deviation",
|
||||
"violence",
|
||||
"self_harm",
|
||||
"doxxing",
|
||||
"scam",
|
||||
"misinformation",
|
||||
"nsfw_image",
|
||||
"gore_image",
|
||||
"illegal_content",
|
||||
"gambling",
|
||||
"drugs",
|
||||
"child_safety",
|
||||
"financial_scam",
|
||||
"religious_insult",
|
||||
"self_promo",
|
||||
]);
|
||||
|
||||
function normalizeFlag(value: string): string | null {
|
||||
const lower = value.trim().toLowerCase().replace(/[\s-]+/g, "_");
|
||||
if (!lower) return null;
|
||||
if (VALID_PRIMARY_AI_FLAGS.has(lower)) return lower;
|
||||
return null;
|
||||
}
|
||||
|
||||
function extractFlagsFromContent(content: string): string[] {
|
||||
const flags = new Set<string>();
|
||||
let parsed: unknown;
|
||||
try {
|
||||
parsed = JSON.parse(content);
|
||||
} catch {
|
||||
parsed = null;
|
||||
}
|
||||
|
||||
const addValue = (v: unknown) => {
|
||||
if (typeof v !== "string") return;
|
||||
const n = normalizeFlag(v);
|
||||
if (n) flags.add(n);
|
||||
};
|
||||
|
||||
if (Array.isArray(parsed)) {
|
||||
for (const item of parsed) addValue(item);
|
||||
} else if (parsed && typeof parsed === "object") {
|
||||
const obj = parsed as Record<string, unknown>;
|
||||
for (const key of ["flags", "categories", "badwords"]) {
|
||||
const val = obj[key];
|
||||
if (Array.isArray(val)) {
|
||||
for (const item of val) addValue(item);
|
||||
} else {
|
||||
addValue(val);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (flags.size > 0) return Array.from(flags);
|
||||
|
||||
const lower = content.toLowerCase();
|
||||
for (const flag of VALID_PRIMARY_AI_FLAGS) {
|
||||
if (lower.includes(flag)) flags.add(flag);
|
||||
}
|
||||
|
||||
return Array.from(flags);
|
||||
}
|
||||
@@ -4,9 +4,9 @@ import { z } from "zod";
|
||||
import { config } from "../config.js";
|
||||
import { createChildLogger } from "../logger.js";
|
||||
import { retryWithBackoff } from "../retry.js";
|
||||
import { withLlmConcurrency } from "./concurrencyLimiter.js";
|
||||
import { resizeImageForVision } from "./imageResizer.js";
|
||||
import { formatModerationTextEvidenceForPrompt } from "./indonesianTextNormalizer.js";
|
||||
import { llmChat, llmVision } from "./llmClient.js";
|
||||
import { extractMessageMediaEvidence } from "./messageMetadata.js";
|
||||
import { buildSystemPrompt as buildSystemPromptModular } from "./moderationPrompt.js";
|
||||
import {
|
||||
|
||||
Reference in New Issue
Block a user